shenao1995/RadGS-Reg-Test

RadGS-Reg — complete test split (431 vertebrae)

The entire fixed RadGS-Reg test split, with 431 lumbar vertebrae from 87 CTs, all 21,550 simulated projections, ground-truth volumes, point clouds, geometry, and the corresponding registration test assets. This supersedes the 10-case demonstration subset as the complete benchmark.

RecM · RegM · Code · Paper

Contents

• Root <sample>.pickle: original float32 128³ attenuation volume, 50 float32 256² projections, angles, scanner parameters and portable metadata. Numerical arrays are unchanged; local source paths in embedded/sidecar metadata become basenames, and file mtimes are omitted.
• <sample>.json and <sample>.ply: matching geometry/provenance and original 10,000-point cloud.
• splits/test.txt: exactly the original 431 filenames in original order.
• registration/manifest.json: only the same 431 test samples, sampled moving-to-fixed GT transforms and physical protocol. No training/validation records are included.
• registration/landmarks/: original fixed physical landmarks for mTRE.
• registration/arrays/: original float16 3-channel [canonical, moving, frozen-RecM] caches, 96³ at 2 mm over a 192 mm canvas.
• registration/rec128/: original float16 frozen-RecM predictions, 128³ at 1 mm.
• registration/splits/test.txt, release.json, array_hashes.json and manifest.json: split, counts, original numerical-array hashes and every published file's SHA256/size.

Frozen caches were generated with the original 2026-10-22 projection-only RecM, SHA256 bcf73bd1415f43cf008a815a7c68d22be27b050b2870454d1e49d465503cf7ef. They are preserved as original test assets, not labeled as predictions from the newer joint model. Joint inference/evaluation uses the CT/pose/landmarks and reconstructs RecM outputs online. The current point RegM operates on 192³ at 1 mm; the older 96³ caches are not its direct training-resolution inputs.

Geometry and protocol

Volumes use XYZ axes in a centered RAS-mm physical frame, 1 mm isotropic spacing. Native projections cover 50 angles in [0°,180°), AP/lateral indices 25/0; the RecM loader area-resizes to 128². Stored projection values are physical line integrals / 1000; volumes use attenuation 1/mm. Use the released loader to preserve scene/input/density scales. Do not replace native inputs with display-normalized thumbnails.

Registration transforms are moving-to-fixed. Pose sampling uses active Rz @ Ry @ Rx, per-axis rotation in [-5°,5°], translation in [-10,10] mm, fixed seed42; matrices are stored for every test sample. CT grouping exactly matches the original split. The complete generation dataset has 3441 training, 431 validation and 431 test vertebrae; this repository publishes test payloads only, with zero CT-group overlap with training/validation.

Level	Test vertebrae
L1	87
L2	87
L3	87
L4	86
L5	84
Total	431

Download and use

from huggingface_hub import snapshot_download
snapshot_download('shenao1995/RadGS-Reg-Test', repo_type='dataset', local_dir='RadGS_test')

Use --data-dir RadGS_test --reg-data-dir RadGS_test/registration --split test for joint inference/evaluation. Use gs_infer.py --data-fold RadGS_test for RecM inference. The repository provides native research binary payloads, not a datasets.load_dataset() builder. Keep each sample's companion files together. Load pickle only from a trusted source; SHA256 manifests are included.

The latest model checkpoints were selected on validation. Uploading these test samples does not constitute a new test evaluation, and no new test performance numbers are claimed here.

Provenance and license

Derived from CTSpine1K, with NanoDRR simulated projections. The source card states CC-BY-NC-SA without a version; preserve attribution, noncommercial and share-alike conditions. See LICENSE.md, SOURCE_CTSpine1K_README.md, the original CTSpine1K project, and the RadGS-Reg paper. This dataset is simulated cropped vertebral research data.